Unmanned aerial vehicle satellite navigation deception jamming detection method and system based on adversarial neural network

Abstract

The application provides an unmanned aerial vehicle satellite navigation deception jamming detection method and system based on an adversarial neural network, and relates to the technical field of unmanned aerial vehicle systems.The method comprises the following steps: calculating the mutual information quantity and time sequence dependence strength between any two feature vectors in an original feature set, and constructing a feature correlation graph with a full connection topology; adopting a multi-head attention mechanism to perform multi-level aggregation and update on the node features in the feature correlation graph, so as to obtain a composite feature tensor that fuses the collaborative change information of each signal element on the time step and the spatial channel.The application improves the detection accuracy and robustness of the unmanned aerial vehicle to satellite navigation deception jamming, and can cope with unknown deception attacks in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) system technology, and in particular to a method and system for detecting UAV satellite navigation deception and interference based on adversarial neural networks. Background Technology

[0002] When drones operate in urban canyon environments, they mostly rely on satellite navigation signals for positioning and flight control. However, progressive synchronization deception jamming, as a highly concealed form of jamming, has become one of the important factors affecting the navigation safety of drones. This deception jamming usually first receives real satellite signals, generates false signals that are synchronized with the real signals in time, and then gradually increases the power of the false signals, causing the drone receiver to smoothly switch from the real signals to the false signals. Subsequently, it slowly adjusts the ranging data of the false signals, inducing the drone to output coordinates that gradually deviate from the real position. The modulation method and pseudocode structure of its false signals are highly similar to those of the real signals, making them difficult to identify by traditional detection methods.

[0003] Currently, detection methods for UAV satellite navigation spoofing interference may have some limitations when dealing with progressive synchronous spoofing interference in urban canyon environments. For example, most detection methods only extract single-dimensional signal features (such as carrier-to-noise ratio) as the detection basis, failing to fully consider the influence of building obstruction and multipath reflection on satellite signals in urban canyon environments. This results in poor stability of single features and makes it difficult to capture the continuous and gradual change pattern of signal features during progressive synchronous spoofing interference. In practical applications, when UAVs fly in densely populated urban areas with tall buildings, progressive synchronous spoofing interference signals will mimic the changing trend of real satellite signals. The changes in features such as carrier-to-noise ratio and code phase offset are relatively gradual, with little difference from the normal fluctuations of real signals affected by urban canyon environments. Existing detection methods, due to their failure to effectively integrate the coordinated change information of multi-dimensional features and the lack of modeling for such gradually changing spoofing signals, may struggle to quickly capture the early characteristics of spoofing interference, leading to detection delays and even the inability to accurately distinguish between normal signal fluctuations and spoofing interference. This could result in UAVs being induced to shift positions by spoofing interference without timely warning, affecting operational safety. Summary of the Invention

[0004] This invention provides a method and system for detecting satellite navigation deception interference by unmanned aerial vehicles (UAVs) based on adversarial neural networks, which improves the detection accuracy and robustness of UAVs against satellite navigation deception interference and helps to cope with unknown deception attacks in complex dynamic environments.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for detecting UAV satellite navigation deception and interference based on adversarial neural networks, the method comprising: Step 1: For the satellite navigation receiver terminal mounted on the belly of the UAV with the antenna phase center located above the fuselage structure, in the multipath reflection micro-environment of urban canyons where the ratio of building height to street width is greater than 2, the satellite navigation baseband signal is collected in real time, and the multi-dimensional signal primitive features of the satellite navigation baseband signal are extracted. The signal primitive features of each dimension are normalized with zero mean and unit variance and nonlinear dimensionality reduction mapping are performed to obtain the corresponding feature vectors. The feature vectors are aligned by sliding window along the time axis and spliced ​​with multi-dimensional tensors to form the original feature set. Step 2: Calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology; use a multi-head attention mechanism to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time step and spatial channel. Step 3: Input the composite feature tensor into the pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork; simulate the continuous gradient offset sequence of progressive synchronous deception interference through the generator subnetwork to obtain the pseudo-deception feature tensor; extract discriminative deep features through forward adversarial discrimination of the discriminator subnetwork based on the pseudo-deception feature tensor and the original composite feature tensor to obtain the discriminative deep features; calculate the deception probability score sequence for the current navigation signal based on the discriminative deep features. Step 4: Compare the deception probability score sequence with the preset detection threshold. If the deception probability score is higher than the detection threshold, it is determined that there is a progressive synchronous deception interference and an alarm signal is generated.

[0006] Secondly, a UAV satellite navigation deception and interference detection system based on adversarial neural networks includes: The feature extraction module is used to collect satellite navigation baseband signals in real time in the multipath reflection micro-environment of urban canyons where the ratio of building height to street width is greater than 2, for satellite navigation receiver terminals mounted on the belly of UAVs with the antenna phase center located above the fuselage structure. The module extracts multi-dimensional signal primitive features of the satellite navigation baseband signals, performs zero-mean unit variance normalization and nonlinear dimensionality reduction mapping on each dimension of the signal primitive features to obtain the corresponding feature vectors, and aligns each feature vector along the time axis by sliding window and splices it with multi-dimensional tensors to form the original feature set. The fusion module is used to calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology. A multi-head attention mechanism is used to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time step and spatial channel. The deception detection module is used to input a composite feature tensor into a pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork. The generator subnetwork simulates a continuous, gradually shifting sequence of progressive synchronous deception interference to obtain a pseudo-deception feature tensor. Based on the pseudo-deception feature tensor and the original composite feature tensor, discriminative deep features are extracted through forward adversarial discrimination by the discriminator subnetwork. Based on the discriminative deep features, a deception probability score sequence for the current navigation signal is calculated. The alarm generation module compares the deception probability score sequence with a preset detection threshold. If the deception probability score is higher than the detection threshold, it determines that there is progressive synchronous deception interference and generates an alarm signal.

[0007] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By extracting satellite navigation signal primitive features in parallel and multi-dimensional ways in the complex environment of urban canyons and constructing an original feature set, a fully connected feature association graph is built based on mutual information and temporal dependency strength. Combined with a multi-head attention mechanism, the spatiotemporal collaborative information of multi-dimensional signal features is deeply aggregated. The intrinsic correlation and temporal evolution law between the carrier-to-noise ratio, code phase shift, Doppler frequency shift and signal angle of arrival of navigation signals are explored. Then, the generator of the adversarial neural network is used to simulate the gradual shift characteristics of progressive synchronous deception interference, and the discriminator completes adversarial discrimination and deep feature extraction and outputs deception probability score. It is adapted to the complex scene characteristics of urban canyons with many obstructions, strong multipath interference and gradual concealment of deception interference, and improves the recognition sensitivity and detection accuracy of progressive synchronous satellite navigation deception interference. It can realize the rapid discrimination and timely warning of interference. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the UAV satellite navigation deception and interference detection method based on adversarial neural networks provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of a UAV satellite navigation deception and interference detection system based on an adversarial neural network, provided by an embodiment of the present invention.

[0011] Figure 3 This is a deception probability scoring curve provided by an embodiment of the present invention.

[0012] Figure 4 This is a comparison chart of smoothing score and detection threshold provided by an embodiment of the present invention. Detailed Implementation

[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0014] like Figure 1 As shown, embodiments of the present invention propose a method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks, the method comprising the following steps: Step 1: For the satellite navigation receiver terminal mounted on the belly of the UAV with the antenna phase center located above the fuselage structure, in the multipath reflection micro-environment of urban canyons where the ratio of building height to street width is greater than 2, the satellite navigation baseband signal is collected in real time, and the multi-dimensional signal primitive features of the satellite navigation baseband signal are extracted. The signal primitive features of each dimension are normalized with zero mean and unit variance and nonlinear dimensionality reduction mapping are performed to obtain the corresponding feature vectors. The feature vectors are aligned by sliding window along the time axis and spliced ​​with multi-dimensional tensors to form the original feature set. Step 2: Calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology; use a multi-head attention mechanism to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time step and spatial channel. Step 3: Input the composite feature tensor into the pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork; simulate the continuous gradient offset sequence of progressive synchronous deception interference through the generator subnetwork to obtain the pseudo-deception feature tensor; extract discriminative deep features through forward adversarial discrimination of the discriminator subnetwork based on the pseudo-deception feature tensor and the original composite feature tensor to obtain the discriminative deep features; calculate the deception probability score sequence for the current navigation signal based on the discriminative deep features. Step 4: Compare the deception probability score sequence with the preset detection threshold. If the deception probability score is higher than the detection threshold, it is determined that there is a progressive synchronous deception interference and an alarm signal is generated.

[0015] In this embodiment of the invention, satellite navigation signal primitive features are extracted in parallel and in multiple dimensions in the complex environment of urban canyons, and an original feature set is constructed. A fully connected feature association graph is built based on mutual information and temporal dependency strength, and a multi-head attention mechanism is combined to complete the deep aggregation of spatiotemporal collaborative information of multi-dimensional signal features. The intrinsic correlation and temporal evolution law between the carrier-to-noise ratio, code phase shift, Doppler frequency shift and signal angle of arrival of navigation signals are explored. Then, the generator of the adversarial neural network is used to simulate the gradual shift characteristics of progressive synchronous deception interference, and the discriminator completes adversarial discrimination and deep feature extraction and outputs deception probability score. This is adapted to the complex scene characteristics of urban canyons with many obstructions, strong multipath interference and gradual concealment of deception interference, and improves the recognition sensitivity and detection accuracy of progressive synchronous satellite navigation deception interference, so as to realize the rapid discrimination and timely alarm of interference.

[0016] In a preferred embodiment of the present invention, step 1 above, for a satellite navigation receiving terminal mounted on the belly of a UAV with its antenna phase center located above the fuselage structure, in a multipath reflection micro-environment of an urban canyon where the ratio of building height to street width is greater than 2, real-time acquisition of satellite navigation baseband signals is performed, and multi-dimensional signal primitive features of the satellite navigation baseband signals are extracted. Zero-mean unit variance normalization and nonlinear dimensionality reduction mapping are performed on each dimension of the signal primitive features to obtain the corresponding feature vectors. The feature vectors are then aligned along the time axis using a sliding window and concatenated with multi-dimensional tensors to form the original feature set, which may include: In this embodiment of the invention, a satellite navigation receiver terminal (selected as a high-precision receiver terminal supporting multiple antenna arrays and adjustable sampling rate, such as a GNSS multi-mode receiver) is fixedly mounted on the belly of the UAV using a shock-resistant bracket. During installation, the terminal's attitude is strictly controlled to ensure that the phase center of the terminal antenna is located above the fuselage structure and that the antenna faces the sky without obstruction. The core purpose of this installation method is to reduce the obstruction and reflection of satellite signals by the UAV's metal structure (such as the arms and fuselage shell), avoiding the superposition of reflected signals and real satellite signals to form false signals, which would lead to distortion in subsequent feature extraction. Simultaneously, the shock-resistant bracket can reduce the impact of UAV vibration during flight on the receiver terminal, ensuring the stability of the signals acquired by the terminal and adapting to complex scenarios in urban canyons with high-rise building obstruction and frequent multipath reflections. The satellite navigation receiver terminal is initialized with parameters, including signal reception frequency bands (covering GPS L1 and BDS). The system initializes the terminal using mainstream satellite navigation frequency bands such as B1C, RF gain (initially set to 40dB, dynamically adjustable based on signal strength), intermediate frequency bandwidth (set to 2MHz, balancing signal acquisition range and anti-interference capability), and data output format (using I / Q baseband signal format for easy feature extraction). After initialization, a terminal self-test program is initiated to verify the terminal's signal reception sensitivity, sampling accuracy, and multi-antenna collaborative operation status, ensuring the terminal is fault-free and can normally acquire satellite navigation signals. The system also utilizes the UAV's onboard environmental perception module (composed of LiDAR, visual sensors, and a geographic information database) to... The system continuously collects building height and street width parameters in the current flight area. It uses lidar to measure the vertical distance between the top of the building and the ground (measurement error controlled within ±0.5m), and uses a visual sensor combined with a geographic information database to calibrate the street width (measurement error controlled within ±0.1m). It calculates the ratio of the two (building height / street width) and selects only areas with a ratio greater than 2 as signal acquisition areas. These areas are urban canyon areas with strong multipath reflection, severe building obstruction, and complex signal propagation paths. The signal acquisition process is only started in these areas to avoid invalid signal acquisition in non-target areas (such as open spaces).

[0017] The antenna array of the satellite navigation receiver terminal receives navigation radio frequency signals from multiple satellites in real time. The received signals first enter the terminal's built-in signal conditioning module, where they undergo three processes in sequence: First, radio frequency filtering, using a bandpass filter (center frequency matched to the receiving frequency band, bandwidth 2MHz) to filter out ground clutter in urban canyons, building reflection interference signals, and electromagnetic interference from other frequency bands (such as urban communication base station signals and industrial electromagnetic interference); Second, low-noise amplification, using a low-noise amplifier (noise figure ≤1.5dB) to amplify the weak radio frequency signal to a processable range (amplified to about 1Vpp) while avoiding signal distortion. Thirdly, down-conversion processing is performed, where the RF signal is converted to an intermediate frequency (IF) signal (IF frequency set to 460MHz) via a mixer, and then further filtered to remove noise, ultimately converting the IF signal into an I / Q baseband signal (in-phase and quadrature components). Considering the signal fluctuation characteristics of the urban canyon environment, an adaptive sampling strategy is adopted. Specific parameter settings and adjustment rules are as follows: the sampling frequency is dynamically adjusted according to the current signal strength. When the signal strength is ≥-120dBm, the sampling frequency is set to 1MHz to ensure real-time signal acquisition; when the signal strength is <-120dBm (affected by building obstruction and multipath reflection), the sampling frequency is adjusted accordingly. When the signal weakens, the sampling frequency is increased to 5MHz to ensure complete capture of the detailed features of the weak signal; the sampling bit depth is set to 16 bits, and the sampling accuracy is controlled within ±0.01V to ensure that the acquired baseband signal can accurately reflect the amplitude and phase changes of the signal. Simultaneously, the acquisition period is linked to the drone's flight speed: when the flight speed is ≤5m / s, the acquisition period is set to 10ms (i.e., 100 data sets are acquired per second); when the flight speed is >5m / s, the acquisition period is shortened to 5ms (i.e., 200 data sets are acquired per second) to avoid discontinuous signal acquisition due to drone movement and to ensure that the acquired baseband signal accurately reflects the signal's characteristics. The current actual status of the navigation signal; the collected I / Q baseband signals are transmitted in real time to the UAV's onboard cache module (using a high-speed cache with a capacity of ≥16GB and a read / write speed of ≥100MB / s). The cache module adopts a circular caching mechanism to retain the collected data of the most recent 10 minutes while avoiding cache overflow. After each set of signal acquisition is completed, the cache module automatically verifies the data. The verification includes data integrity (whether there are missing values) and data validity (whether there are abnormal sudden values, such as a sudden spike or drop in signal amplitude). If data abnormality is found, re-acquisition is immediately triggered to ensure that the collected baseband signals are complete and without any abnormalities.

[0018] The carrier-to-noise ratio (CNR) time series reflects the strength and purity of satellite signals and is a fundamental characteristic for determining whether a signal is interfered with. It can capture the gradual changes in signal strength during progressive spoofing interference. The extraction algorithm uses the coherent integration method. Specifically, for each group of I / Q baseband signals, coherent integration is performed for 1ms, and the ratio of the square of the integrated signal amplitude to the noise power is calculated to obtain the CNR at a single time point (in dB-Hz). The sampling interval is kept consistent with the signal acquisition period (5ms or 10ms), and the CNR time series is obtained by continuous acquisition. To resist the influence of multipath reflections in urban canyons, the integrated signal is denoised to ensure that the CNR calculation error is ≤0.5dB-Hz. After each baseband signal acquisition is completed, a CNR value is output synchronously to form a continuous CNR time series.

[0019] Code phase offset reflects the synchronization state between the satellite signal pseudocode and the local pseudocode of the receiving terminal. Multipath reflections in urban canyons can cause slight code phase offsets (the offset is usually ≤0.1 chip), while progressive spoofing interference can cause the code phase to show a continuous and gradual offset (the offset gradually increases, and the initial difference from normal fluctuations is minimal). This subtle difference needs to be captured through this feature. The extraction algorithm uses a delay-locked loop (DLL). Specifically, the received I / Q baseband signal is correlated with the pseudocode generated locally by the terminal. The delay of the local pseudocode is adjusted to maximize the correlation value. At this point, the delay difference between the local pseudocode and the received signal pseudocode is the code phase offset (unit: chip). The loop bandwidth of the delay-locked loop is set to 10Hz, and the tracking accuracy is controlled within ±0.01 chip to ensure that minute code phase changes can be captured. After each set of baseband signal acquisition is completed, a code phase offset value is output synchronously, forming a continuous code phase offset sequence.

[0020] Doppler frequency shift residuals reflect the deviation between the satellite signal frequency and the theoretical frequency. They can resist frequency fluctuation interference caused by drone movement and building obstruction in urban canyons, and can also capture the gradual frequency drift in progressive deception interference. The extraction process consists of two steps. First, based on the flight speed and attitude obtained by the drone's onboard GPS / IMU integrated navigation module, combined with satellite orbit parameters (obtained in real time from the satellite navigation system), the theoretical Doppler frequency shift is estimated using the Doppler frequency shift calculation formula. Second, a phase-locked loop (PLL) is used to perform frequency tracking on the I / Q baseband signal to calculate the actual Doppler frequency shift. The loop bandwidth of the PLL is set to 5Hz, and the frequency tracking accuracy is controlled within ±1Hz. The actual Doppler frequency shift is subtracted from the theoretical Doppler frequency shift to obtain the Doppler frequency shift residual (in Hz). After each set of baseband signal acquisition is completed, a Doppler frequency shift residual value is output synchronously, forming a continuous Doppler frequency shift residual sequence.

[0021] The signal angle of arrival (AHA) estimate reflects the spatial origin of the satellite signal, distinguishing the spatial differences between real satellite signals and multipath reflection signals and spoofing interference signals. It compensates for the limitations of single time-domain and frequency-domain features, and is suitable for scenarios in urban canyons where multipath reflection signals and spoofing signals are mixed. The extraction algorithm employs the MUSIC (Multiple Signal Classification) algorithm, relying on a multi-antenna array at the receiving terminal (using a 4-element antenna with an element spacing of 0.5λ, where λ is the satellite signal wavelength). Specifically, the covariance matrix of the I / Q baseband signals acquired by the multi-antenna array is estimated. Eigenvalue decomposition separates the signal subspace from the noise subspace. Utilizing the orthogonality between the noise subspace and the AHA, the azimuth and elevation angles of the satellite signal are calculated, yielding the estimated AHA (azimuth range: 0°-360°, elevation range: 0°-90°). The estimation accuracy is controlled within ±1°. For each set of baseband signal acquisitions completed, a signal AHA estimate is output synchronously, forming a continuous sequence of estimated AHAs.

[0022] The extraction of signal primitive features in the above four dimensions is carried out simultaneously using a parallel processing architecture. After each baseband signal acquisition (5ms or 10ms), a set of four-dimensional feature data (1 carrier-to-noise ratio value, 1 code phase offset value, 1 Doppler frequency shift residual value, and 1 signal angle of arrival estimate) is extracted to ensure that the features are synchronized with the signal acquisition in real time and to avoid time misalignment caused by feature extraction lag.

[0023] Since the extracted multi-dimensional signal primitive features belong to different physical quantities with large differences in numerical range (e.g., carrier-to-noise ratio range: 30-50dB-Hz, code phase offset range: 0-0.5 chip, Doppler frequency shift residual range: -10-10Hz, signal angle of arrival estimation range: 0°-360°), directly using them for subsequent feature association and fusion will lead to an imbalance in feature weights (e.g., the numerical range of the signal angle of arrival is much larger than that of the carrier-to-noise ratio, which will dominate subsequent calculations), affecting the accuracy of the processing results. At the same time, the feature data collected in the urban canyon environment contains certain redundant information (e.g., the correlation between adjacent data in the carrier-to-noise ratio time series is high), which will increase the computational load of subsequent feature association graph construction and adversarial neural network processing, reducing real-time performance. Therefore, zero-mean unit variance normalization and nonlinear dimensionality reduction mapping need to be performed on the signal primitive features of each dimension. The specific process, algorithm details, and parameter settings are as follows. Zero-mean unit variance normalization (Standardization) involves independently standardizing each signal primitive feature (carrier-to-noise ratio time series, code phase offset series, Doppler shift residual series, and signal angle of arrival estimation series) to eliminate dimensional differences between features and ensure equal weighting for each feature. The specific steps include: data statistics, calculating the mean μ and standard deviation σ of all sampled data for a given feature series, where μ is the arithmetic mean of all data in that dimension, and σ is the square root of the sum of squares of the differences between all data in that dimension and the mean. A sliding window (100 data sets) is used to update μ and σ in real time, adapting to the dynamic changes in signal characteristics in urban canyon environments and avoiding standardization distortion caused by fixed statistical values; and normalization calculation using the formula... Each feature data point in this dimension is transformed into standardized data with a mean of 0 and a variance of 1. The original feature data, For the standardized feature data, for example, if the original carrier-to-noise ratio (CNR) data is 40 dB-Hz, the mean μ = 38 dB-Hz, and the standard deviation σ = 2 dB-Hz, then the standardized data is (40-38) / 2 = 1; if the original code phase offset data is 0.2 chip, the mean μ = 0.1 chip, and the standard deviation σ = 0.05 chip, then the standardized data is (0.2-0.1) / 0.05 = 2; for outlier handling, after standardization, if the absolute value of a certain feature data is greater than 3 (i.e., exceeds 3 times the standard deviation), it is judged as an outlier (possibly caused by extreme interference or data acquisition errors). The outlier is replaced by linear interpolation (using the average of two adjacent normal data) to ensure that the standardized feature sequence is free of outliers and avoids affecting subsequent dimensionality reduction processing.

[0024] Nonlinear dimensionality reduction mapping employs the t-SNE (t-distributed neighborhood embedding) algorithm for nonlinear dimensionality reduction of each dimension of the normalized feature sequence. This algorithm effectively removes redundant information while preserving the core variation patterns of features, and avoids the shortcomings of linear dimensionality reduction algorithms (such as PCA) in capturing nonlinear feature correlations. It is suitable for nonlinear changes in signal features in urban canyons (such as feature abrupt changes caused by multipath reflections and gradual changes caused by deception interference). Specific operation steps and parameter settings include setting the perplexity of the t-SNE algorithm to 30 (reflecting the size of the local neighborhood of the feature data, suitable for feature sequence lengths of 100-200 sets), the learning rate to 200, and the number of iterations to 1000 to ensure the convergence of the dimensionality reduction process; the dimension of the feature vector after dimensionality reduction is uniformly set to 128 dimensions, which can both preserve the core information of the features and significantly reduce the subsequent computational load (the original feature sequence has a dimension of 100 dimensions for every 100 sets of data, which is compressed to 128 dimensions after dimensionality reduction, balancing information preservation and computational efficiency); the dimensionality reduction process involves modifying the normalized feature sequence of a certain dimension... The algorithm takes a column (such as a carrier-to-noise ratio time series) as input and calculates the coordinates of each feature data in the low-dimensional space using the t-SNE algorithm, forming a 128-dimensional feature vector. During this process, the algorithm focuses on preserving the temporal evolution of the feature sequence (such as the continuous rise and fall trend of the carrier-to-noise ratio and the gradual shift of the code phase), removes redundant correlations between adjacent data, and strengthens the abnormal changes of features (such as the weak gradual changes in the early stage of deception interference), laying the foundation for subsequent capture of deception interference features. For effect verification, after dimensionality reduction, the correlation coefficient of the feature sequence before and after dimensionality reduction is calculated to ensure that the correlation coefficient is ≥0.95, that is, the feature vector after dimensionality reduction can retain more than 95% of the core information of the original feature sequence. If the correlation coefficient is <0.95, the perplexity parameter of the t-SNE algorithm is adjusted (adjustment range 20-40), and dimensionality reduction is performed again until the requirements are met. After the above processing, the signal primitive features of each dimension are converted into 128-dimensional standardized feature vectors. The four dimensions correspond to four 128-dimensional feature vectors, and the dimensions of each vector are uniform, the units are consistent, and there is no redundant information, providing good conditions for sliding window alignment and multi-dimensional tensor splicing.

[0025] The sliding window parameters are set reasonably, taking into account the UAV's flight speed, signal acquisition cycle, and the characteristic change cycle of deception interference. This ensures that the window can continuously cover all feature data while balancing temporal resolution and feature correlation. Specifically, the window length is determined based on the signal acquisition cycle. When the acquisition cycle is 5ms, the window length is set to 20 acquisition cycles (i.e., 100ms); when the acquisition cycle is 10ms, the window length is set to 15 acquisition cycles (i.e., 150ms). This length can capture the coordinated changes of multi-dimensional features within a short period (100-150ms) while avoiding the problems caused by an excessively long window. The feature lag is adapted to the subtle feature changes in the early stages of gradual deception interference (the initial feature changes are small and require accumulation through a window of a certain length to capture); the sliding step size is set to one acquisition cycle (5ms or 10ms), meaning that after each acquisition, the window slides forward one acquisition cycle, ensuring that the window can continuously cover all feature data without omission, while avoiding feature loss due to excessively large sliding steps, ensuring that the original feature set can completely reflect the temporal evolution of the signal features; the sliding window is aligned, aligning the four-dimensional feature vectors (carrier-to-noise ratio feature vector, code phase offset feature vector, Doppler frequency shift residual feature vector, signal...). The angle of arrival (Angle of Arrival) estimation feature vector is used as a time reference, and strict sliding window alignment is performed to ensure that multi-dimensional feature data within the same window correspond to satellite navigation signals within the same time period, avoiding feature association distortion caused by time misalignment. Specifically, timestamp synchronization involves adding a timestamp (with a precision of 1ms) to each data point in the feature vector for each dimension. The timestamps are kept consistent with the signal acquisition time, ensuring complete synchronization of the four-dimensional feature data timestamps (error ≤ 1ms). Window data extraction involves extracting four features synchronously based on the timestamps for each sliding window (e.g., the first window, with a time range of 0-100ms). For feature data with timestamps in the dimensional feature vectors ranging from 0 to 100ms, 20 data points are extracted from each dimension (corresponding to 20 acquisition cycles), forming four 128×20 feature matrices (each data point is a 128-dimensional feature vector). Alignment verification is performed: after each window alignment is completed, the consistency of timestamps in the four dimensional feature matrices is verified. If the timestamp of a feature data in a certain dimension deviates from the window time range by more than 1ms, it is considered an alignment failure, and the feature data of that dimension is re-extracted to ensure that the time synchronization of multi-dimensional feature data within the same window provides an accurate time synchronization basis for calculating the mutual information between features and constructing feature association graphs.

[0026] Multidimensional tensor concatenation involves concatenating the aligned four-dimensional feature matrices within each sliding window along the feature channel dimension to construct the feature tensor corresponding to a single window. The specific concatenation logic and tensor dimension definition result in a three-dimensional tensor. The three dimensions are defined as follows: time dimension (corresponding to the number of acquisition cycles within the window, i.e., 20 or 15), feature dimension (corresponding to the dimension of the reduced feature vector, i.e., 128 dimensions), and feature channel dimension (corresponding to the four signal primitive features, i.e., 4 channels). The tensor dimension is represented as [time dimension × feature dimension × channel dimension]. For example, if the acquisition cycle is 5m... When the window length is 20, the tensor dimension is [20×128×4]. During the splicing process, the four-dimensional feature matrices (128-dimensional × 20) are spliced ​​sequentially according to the channel order. That is, the first channel is the carrier-to-noise ratio feature matrix, the second channel is the code phase offset feature matrix, the third channel is the Doppler frequency shift residual feature matrix, and the fourth channel is the signal angle of arrival estimation feature matrix. During the splicing process, it is ensured that the time dimension and feature dimension of each matrix are completely aligned without misalignment. The three-dimensional tensor formed after splicing completely preserves the temporal evolution law and spatial coordination relationship of the multi-dimensional features within a single window.

[0027] The initial feature set is constructed and validated by arranging the 3D feature tensors corresponding to all sliding windows in chronological order. For example, if a drone flies over an urban canyon area for 10 minutes with a data acquisition period of 5ms, collecting 120,000 sets of signal data, and the sliding window length is 20 with a sliding step size of 1, a total of 119,981 sliding window tensors are formed. Arranging these tensors in chronological order constitutes the initial feature set. The initial feature set is validated in two ways: first, completeness verification, ensuring that all sliding window tensors are free of missing or abnormal features (such as incorrect tensor dimensions or empty data), with a 100% pass rate; second, correlation verification, calculating the correlation coefficients of the four channel features within the same window to ensure that each channel... There are reasonable correlations between the features (such as the correlation coefficient between carrier-to-noise ratio and code phase offset being between 0.3 and 0.5, which conforms to the correlation law of signal features in urban canyon environment). If the correlation coefficient is abnormal (such as <0.1 or >0.8), the feature extraction, standardization or alignment process is re-examined to ensure the quality of the original feature set. The final original feature set fully preserves the temporal evolution law and spatial coordination relationship of multi-dimensional signal primitive features, while eliminating environmental interference and feature redundancy. It can provide high-quality input data for calculating the mutual information between features and constructing feature correlation graphs, laying a solid foundation for the early detection of progressive synchronous deception interference. It is fully adaptable to the multipath reflection microenvironment of urban canyons and the actual flight operation scenario of UAVs.

[0028] In a preferred embodiment of the present invention, step 2 above, which calculates the mutual information and temporal dependency strength between any two feature vectors in the original feature set, constructs a feature association graph with a fully connected topology; and uses a multi-head attention mechanism to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in the time step and spatial channel, may include: In this embodiment of the invention, step 220 involves extracting any two feature vectors from the original feature set, and discretizing and binning the sampling points of these two feature vectors within the same time window. The joint occurrence frequency and the single-dimensional occurrence frequency within each bin are then calculated. Based on the ratio of the joint frequency to the single-dimensional frequency, the mutual information between the two feature vectors is calculated to obtain a first intermediate parameter reflecting the strength of their statistical correlation. Specifically, this includes: the original feature set is... ,in For the first The combined feature vector of time windows has a dimension of 4L ( (length of the sliding time window) Let be the number of time windows, and let any two feature vectors be denoted as . and ( , From 1 to The specific implementation process is as follows: extract any two feature vectors from the original feature set (i ≠ j). and Both feature vectors are 4L-dimensional column vectors, corresponding to the _i_th ... The and the first The multi-dimensional combination features of each time window include normalized feature information in four dimensions: carrier-to-noise ratio, code phase offset, Doppler frequency shift residual, and signal angle of arrival estimate. Since the two feature vectors correspond to signal features in different time windows, matching is required for sampling points within the same time window. Here, sampling points refer to elements within the feature vectors, i.e. The element and The element ( =1 to 4L) correspond to the same positional features of the same signal primitive, ensuring the targeted binning process; the sampling points of these two feature vectors within the same time window are discretized and binned. The purpose of binning is to convert continuous normalized feature values ​​into discrete categories, which facilitates the statistical analysis of joint occurrence frequency and single-dimensional occurrence frequency. The equal-width binning method is used for discretization. The number of bins B is determined according to the dimension and numerical distribution of the feature vectors, usually taking a value of 5 to 10, to ensure that there are enough sampling points in each bin, while avoiding too many bins leading to sparse frequency statistics and too few bins leading to information loss; the specific binning process is as follows: first, determine the numerical range of all sampling points of the two feature vectors. Since the feature vectors have been normalized, the numerical range is [0, 1]. Therefore, the [0, 1] interval is divided into B equal-width sub-intervals, each of which is a bin. The range of each box is (b=1 to B), Each sampling point Assign them to the corresponding bins in the corresponding intervals, and similarly... Each sampling point Assign boxes to the corresponding intervals to ensure and Each element is consistently assigned to its corresponding bin, laying the foundation for subsequent joint frequency statistics; the joint occurrence frequency within each bin and the single-dimensional occurrence frequency within each bin are calculated. The joint occurrence frequency refers to... The sampling point falls into the b1th box and The number of times the corresponding sampling point falls into the b2th bin is denoted as . Where b1 and b2 are integers from 1 to B, and the statistics need to be performed one by one. and Matching and statistical analysis were performed on 4L sampling points, that is, for each (1 to 4L), determine The box b1 and The box b2 where it is located will The value is increased by 1, resulting in a B×B joint frequency matrix; single-dimensional frequency occurrences are divided into two categories, one of which is... The total number of times the sampling point falls into the b1-th bin is denoted as . The calculation formula is: The other type is The total number of times the sampling point falls into the b2th box is denoted as . The calculation formula is: Two one-dimensional vectors, each with a frequency of 1×B, respectively reflect... and Its own numerical distribution characteristics.

[0029] The mutual information between two feature vectors is calculated based on the ratio of joint frequency to single-dimensional frequency. Mutual information measures the statistical correlation between two feature vectors; a higher value indicates a stronger correlation. The formula for calculating mutual information is as follows: ,in Representing the eigenvector and The mutual information between them, i.e. the first intermediate parameter, takes a value in the range of [0, log2B]. Denotes the joint probability, i.e. It fell into the b1 box and The probability of falling into the b2th box is calculated using the following formula: The denominator 4L represents the total number of sampling points for a single feature vector; express The marginal probability of falling into the b1-th box is calculated using the following formula: ; express The marginal probability of falling into the b2th box is calculated using the following formula: log2 is a base-2 logarithmic operation, ensuring that the unit of mutual information is bits. or Then the logarithm of the corresponding term is set to 0 to avoid meaningless values ​​during the calculation, thus obtaining the mutual information. Then, it was used as a reflection and The first intermediate parameter for statistical correlation strength.

[0030] Step 221: Based on the first intermediate parameter and the value change sequence of the same pair of feature vectors between adjacent time windows, calculate the cross-covariance function of the two feature vector sequences on the time axis, and normalize the maximum value of the cross-covariance function within a specified maximum time delay range to obtain the temporal dependency strength reflecting the degree of mutual following change between the two in continuous time steps, which is used as the second intermediate parameter; specifically, this includes: obtaining the value change sequence of the same pair of feature vectors, where the same pair of feature vectors refers to the sequence selected in step 220. and The corresponding sequence of feature vectors, i.e. The feature vector sequence it belongs to is (The original feature set itself) For an element in this sequence, it is expanded into two complete feature vector sequences, denoted as follows: and (Actually, it's the same original feature set, used to simulate the temporal changes of any two feature vector sequences.) The processing logic is consistent for any two different feature vector sequences. The value change sequence refers to the difference sequence of feature vectors between adjacent time windows, used to reflect the trend of feature vector changes over time. For feature vector sequences... Its value change sequence is denoted as ,in ( =1 to K-1), It is a 4L-dimensional column vector, where each element represents the change in the feature at the corresponding location within adjacent time windows; similarly, the feature vector sequence The sequence of value changes is denoted as ,and Consistent (because) and If they are the same sequence, and if they are different feature vector sequences, calculate the value change sequence using the same method.

[0031] Calculate the cross-covariance function of two eigenvector sequences on the time axis. The cross-covariance function measures the linear correlation between two time series at different time delays, reflecting the synergy of the two eigenvector sequences over time. This represents the time offset between two sequences, with a value range of [0, ..., ...]. ],in The maximum specified time delay is determined based on the number of time windows K in the original feature set, typically ranging from K / 5 to K / 3. This ensures coverage of the main temporal correlation range of the sequence while avoiding excessive time delays that could lead to computational redundancy. The formula for calculating the cross-covariance function is as follows: ,in Indicates a time delay of Cross-covariance values ​​at time; This represents the time delay, with values ​​of 0, 1, ... ; express The Elements (i.e.) In the and +1 time window of change). express The + Elements (i.e.) In the + and + +1 time window of change). express The mean vector is calculated using the following formula: ; express The mean vector is calculated using the following formula: ; The denominator represents the transpose operation of a vector; To effectively calculate the sample size and ensure the stability of the cross-covariance, since and Both are 4L-dimensional column vectors, therefore Given a 4L×4L matrix, the trace of this matrix (the sum of the elements on the main diagonal) is taken as the time delay. The scalar value of the cross-covariance at time is denoted as This is used for subsequent normalization; the cross-covariance function is normalized to its maximum value within a specified maximum time delay range to obtain the time dependency strength. First, all time delays are calculated. (0≤) ≤ The corresponding cross-covariance scalar value Find the maximum value among them. ;like If the value is 0, it indicates that the two feature vector sequences have no temporal correlation, and the temporal dependency strength is set to 0; otherwise, the maximum value normalization method is used to normalize each time delay. corresponding Mapping to the [0, 1] interval, the normalization calculation formula is as follows: ,in Representing the eigenvector and The temporal dependency strength between the two features, i.e. the second intermediate parameter, has a value range of [0, 1]. The larger the value, the stronger the mutual following of the two feature vectors at continuous time steps.

[0032] Step 222: Multiply the first intermediate parameter by a preset first weight coefficient, multiply the second intermediate parameter by a preset second weight coefficient, and then add the two products to obtain a comprehensive correlation score. A nonlinear compression transformation is then used to map the comprehensive correlation score to a preset intensity range to obtain the standardized correlation weight between each pair of feature vectors. Specifically, this includes setting preset first and second weight coefficients. These two weight coefficients are used to allocate the proportion of the first and second intermediate parameters in the comprehensive correlation score. They are determined based on the characteristics of the satellite navigation signal and interference detection requirements, using a combination of empirical values ​​and training optimization to ensure that the comprehensive correlation score fully reflects the correlation degree between the two feature vectors. The first weight coefficient is denoted as... The first intermediate parameter (mutual information) has a value range of [0.4, 0.6]; the second weighting coefficient is denoted as... The corresponding second intermediate parameter (temporal dependency strength) has a value range of [0.4, 0.6], and satisfies the following conditions: ,For example , To ensure a balanced weight distribution between the two parameters, considering both statistical correlation and temporal dependence, the comprehensive association score is calculated by multiplying the first intermediate parameter by the first weight coefficient, the second intermediate parameter by the second weight coefficient, and then summing the two products. The formula is as follows: ,in Representing the eigenvector and A comprehensive correlation score between them; The first weighting coefficient; This is the first intermediate parameter (mutual information). This is the second weighting coefficient; The second intermediate parameter (temporal dependency strength); the comprehensive association score ranges from [0, ... The larger the value, the stronger the correlation between the two feature vectors.

[0033] A nonlinear compression transformation is used to map the comprehensive association score to a preset intensity range, resulting in standardized association weights. The preset intensity range is [0, 1], consistent with the numerical range of the normalized features mentioned earlier, facilitating unified weight processing in the subsequent feature association graph. The sigmoid function is used as the nonlinear compression transformation function. This function can map any real number to the [0, 1] interval and has smooth gradient characteristics, making it suitable for compressing association scores. The transformation calculation formula is as follows: ,in Representing the eigenvector and The standardized association weights between them, with values ​​ranging from [0, 1]; This is the compression factor, used to adjust the steepness of the compression curve. It is usually set to 2 to 5 to ensure that the differences in the correlation scores can be effectively reflected. This is the offset coefficient, used to adjust the midpoint position of the compression curve. It is usually taken as the mean of the comprehensive correlation score to ensure that the weight distribution is uniform after compression. is a natural constant, with a value of approximately 2.71828; To achieve a comprehensive correlation score, this nonlinear compression transformation maps all comprehensive correlation scores to the interval [0, 1], obtaining the standardized correlation weight between each pair of feature vectors. The stronger the correlation between two feature vectors, the closer the standardized correlation weight is to 1, and vice versa. This weight will serve as the core basis for the edge weight of the feature correlation graph in step 223.

[0034] Step 223: Based on the standardized association weights and all feature vectors in the original feature set, each feature vector is treated as a graph node, and the standardized association weight between each pair of nodes is used as the weight of that edge. Connection edges are established between all node pairs to construct a feature association graph with a fully connected topology. Specifically, this includes: determining the node set of the feature association graph, the original feature set... Includes There are 3 combined feature vectors, each corresponding to a graph node. Therefore, the node set of the feature association graph is 1. The total number of nodes is Each node has a feature dimension of 4L, and the feature value of the node is the element value of the corresponding combined feature vector, fully preserving the multi-dimensional signal primitive information within the time window; each node corresponds to a unique identifier, and the identifier number is consistent with the time window number, i.e., node 1 corresponds to... (First time window), node 2 corresponds to (Second time window), and so on, node K corresponds to (The Kth time window) facilitates the location and aggregation of subsequent node features; it determines the edge set and edge weights of the feature association graph, which adopts a fully connected topology, meaning that an undirected edge is established between any two nodes, and there are no isolated nodes, ensuring that the feature information of all nodes can be mutually transmitted and fused. For any two nodes... and ( ≠ The weights of the connecting edges between them are the standardized association weights calculated in step 222. For a node itself (i=j), there are no connecting edges, or the edge weight is set to 0 (not participating in subsequent feature aggregation calculation). The edge weight ranges from [0, 1]. The larger the edge weight, the stronger the correlation between the feature vectors of the two nodes, and the higher the weight of the feature information passed between them during feature aggregation. The smaller the edge weight, the weaker the correlation between the two nodes, and the lower the weight of the feature information passed.

[0035] Construct a complete feature association graph by combining the node set and edge set according to a fully connected topology to form an undirected fully connected graph, denoted as . ,in A set of nodes, containing There are 10 nodes, and each node corresponds to a combined feature vector. Let be the set of edges, containing There are undirected edges, and there is an edge between any two distinct nodes; It is a set of edge weights, containing the standardized association weights corresponding to each edge. This feature association graph fully preserves the multi-dimensional signal information of all feature vectors in the original feature set. At the same time, it reflects the statistical correlation and temporal dependence between any two feature vectors through edge weights, connecting the scattered node features through association weights, and providing a structured input carrier for the feature aggregation of the multi-head attention mechanism in step 224.

[0036] Step 224 involves mapping the feature vector of each node in the feature association graph to a query vector, a key vector, and a value vector, respectively. It also calculates the similarity score between the current node and all other nodes' query and key vectors under each attention head, and normalizes the similarity scores to obtain the attention weights between the current node and its neighboring nodes. Specifically, this includes setting the core parameters of the multi-head attention mechanism, including the number of attention heads H and the mapping dimension d. These two parameters are determined based on the node feature dimension (4L) and the required interference detection accuracy, ensuring that the mapped vectors can effectively capture the association information of node features while avoiding computational redundancy. The number of attention heads H is typically between 4 and 8, and each attention head learns independently. Different feature association patterns ultimately improve the feature aggregation effect by fusing the results of multiple attention heads. The mapping dimension d satisfies H×d=4L, meaning each attention head maps a 4L-dimensional node feature vector to a d-dimensional query vector, key vector, and value vector, ensuring that the total dimension after mapping is consistent with the original node feature dimension. For example, when 4L=80 and H=4, d=20. The feature vector of each node is mapped to the query vector, key vector, and value vector respectively. The mapping process is implemented through three independent fully connected layers. The parameters of the three fully connected layers are independent and do not share weights, ensuring that the query vector, key vector, and value vector can capture different dimensions of node features. For any node in the feature association graph... (k=1~K), whose feature vector is a 4L-dimensional column vector, is mapped to a query vector through the first fully connected layer. It is mapped to a key vector through a second fully connected layer. It is mapped to a value vector through the third fully connected layer. The three vectors all have dimension d, and the formula for calculating the fully connected layer is: ,in These are the weight matrices for the three fully connected layers, each with a dimension of d×4L; These are the bias vectors for the three fully connected layers, each with dimension d; They are nodes The query vector, key vector, and value vector are all of dimension d, and the query vector, key vector, and value vector of all nodes respectively constitute the query matrix. Key matrix Value matrix The dimensions are all K×d, where The kth row is , The kth row is , The kth row is Calculate the similarity score between the current node and the query vector and key vector of all other nodes under each attention head. The similarity score measures the degree of feature association between two nodes. The dot product similarity method is used for calculation, and the formula is as follows: ,in This represents the similarity score between the query vector of the k-th node and the key vector of the m-th node (m=1 to K). Let the key vectors be d-dimensional column vectors. The dot product result is a scalar value; the larger the value, the higher the similarity between the two vectors and the closer the node association. To avoid excessively high similarity scores causing gradient vanishing after softmax normalization, the similarity score is divided by 1 / 2. After scaling, the similarity score is: ,in The similarity score is normalized using a scaling factor to obtain the attention weights between the current node and its neighbors. The softmax function is used for normalization to ensure that the sum of the normalized attention weights is 1, facilitating subsequent weighted summation of the value vectors. The normalization formula is as follows: ,in Indicates the first The node and the first The attention weights between nodes range from [0, 1]. ; It is a natural constant; The scaled similarity score is used; the denominator is the exponential sum of the scaled similarity scores of all nodes with the k-th node, used for normalization. For each attention head, the above mapping, similarity calculation, and normalization operations are performed to obtain the attention weight matrix for each attention head. (h=1 to H), all with dimensions K×K, where The element in the k-th row and m-th column is the attention weight between the k-th node and the m-th node under the h-th attention head. .

[0037] Step 225: Weight the value vectors of all nodes according to the attention weights to obtain the single-head aggregated feature of the current node under that attention head. Concatenate or average the single-head aggregated features output by all attention heads for each node, and perform a non-linear transformation to obtain the updated node features of the first-layer graph attention network. Specifically, this includes: calculating the single-head aggregated feature of the current node under each attention head. The single-head aggregated feature is obtained by weighting the value vectors of all nodes according to the attention weights, which can fuse the feature information of the current node and all neighboring nodes. The higher the attention weight of a neighboring node, the greater its contribution to the aggregated feature of the current node. For the h-th attention head, the single-head aggregated feature of the k-th node... The calculation formula is: ,in Indicates the first The single-head aggregated feature of the kth node under each attention head, with dimension d; Indicates the first Attention weights between the k-th node and the m-th node under each attention head; This represents the value vector of the m-th node, with dimension d. The summation operation covers all K nodes, ensuring the fusion of feature information from all nodes. Under each attention head, the single-head aggregated features of all K nodes constitute a single-head aggregated feature matrix. The dimension is K×d, where The kth row is The single-head aggregated features from all attention heads are fused. The fusion method is either concatenation or averaging, chosen based on the feature dimension and interference detection requirements. When the number of attention heads is small and the mapping dimension d is small, concatenation is used to retain more feature details. When the number of attention heads is large and the mapping dimension d is large, averaging is used to reduce feature redundancy and improve computational efficiency. The fusion calculation formula for the concatenation method is as follows: ,in The aggregated feature after fusing the k-th node has a dimension of H×d=4L, consistent with the original node feature dimension; the fusion calculation formula for the averaging method is as follows: ,in The dimension is still d, and it needs to be mapped back to 4L dimension through subsequent fully connected layers. A concatenation method is used to ensure that the feature information of all attention heads is retained. The aggregated features of all K nodes after fusion constitute the fused feature matrix Z, with a dimension of K×4L. A nonlinear transformation is performed on the fused aggregated features to obtain the updated node features of the first layer graph attention network. The nonlinear transformation is used to introduce nonlinear correlations, improve the expressive power of the features, and better capture the collaborative change information of multi-dimensional signals. The nonlinear transformation is implemented through a fully connected layer and an activation function. The weight matrix of the fully connected layer... The dimension is 4L×4L, and the bias vector is... With a dimension of 4L, the ReLU activation function effectively alleviates the gradient vanishing problem and improves feature learning ability. The formula for calculating the nonlinear transformation is as follows: ,in This represents the feature vector of the k-th node after the update of the first-layer graph attention network, with a dimension of 4L; This is the weight matrix of the fully connected layer; The aggregated features are those obtained after fusion; Here is the bias vector of the fully connected layer; the ReLU function is a non-linear activation function, calculated as follows: It can set negative values ​​to 0 and retain the feature information of positive values. The updated feature vectors of all K nodes constitute the first layer update feature matrix U, with a dimension of K×4L. This matrix is ​​the output of the first layer graph attention network.

[0038] Step 226: Based on the node features updated by the first-layer graph attention network, repeatedly perform node feature aggregation and update operations for the multi-layer graph attention network. The input to each layer is the node features output by the previous layer, and each layer independently learns different attention parameters. This allows the final features of each node to fuse the collaborative change information of all nodes in the multi-hop neighborhood, resulting in all node features after multi-layer aggregation and updating. Specifically, this includes: setting the total number of layers in the graph attention network. The total number of layers is based on the number of nodes in the feature association graph. The required interference detection accuracy is determined by the number of layers, typically ranging from 3 to 5. Too few layers will result in insufficient feature aggregation and an inability to capture collaborative information from multi-hop neighborhoods; too many layers will lead to overfitting and increased computational complexity. The structure of each layer of the graph attention network is consistent with the single-layer structure described in steps 224 and 225, both including attention weight calculation, single-head aggregation feature fusion, and nonlinear transformation. However, the parameters of the fully connected layer in each layer ( ) and bias vector ( Each layer learns independently without sharing parameters, ensuring that each layer can learn different feature association patterns and capture collaborative change information at different levels. A multi-layer aggregation update operation is performed. The input to the first layer is the first-layer update feature matrix U (dimension K×4L) obtained in step 225. This matrix is ​​used as the input node features of the second-layer graph attention network. Steps 224 and 225 are repeated, mapping the feature vector of each node to a query vector, key vector, and value vector. The attention weight of each attention head is calculated, and a weighted sum is obtained to obtain the single-head aggregated feature. The results of all attention heads are fused and subjected to a non-linear transformation to obtain the updated node feature matrix of the second-layer graph attention network. (Dimension K×4L), will The above operation is repeated as input to the third-layer graph attention network to obtain the updated node feature matrix of the third layer. And so on, until completion. The aggregation update operation of the layered graph attention network; in the multi-layer aggregation update process, the input of each layer is the node feature matrix output by the previous layer, and the attention parameters of each layer are learned independently, so that the features of each node can gradually fuse the collaborative change information of all nodes in the multi-hop neighborhood. For example, the first layer aggregation only fuses the feature information of the node's direct neighbors (1-hop neighborhood), the second layer aggregation fuses the feature information of the neighbor's neighbor (2-hop neighborhood), and so on, the third layer aggregation... Layer aggregation and fusion By jumping to the feature information of all nodes in the neighborhood, it ensures that the final feature of each node can comprehensively reflect the multi-dimensional signal collaborative change information in the entire feature association graph, effectively capturing the subtle gradual change pattern of signal features caused by progressive synchronous deception interference; and completing... After the aggregation and update of the layered graph attention network, the feature matrix of all nodes after multiple aggregation and update is obtained. (Dimension K×4L), the k-th row of this matrix is ​​the feature vector of the k-th node after multi-level aggregation and updating. (Dimension 4L) Each feature vector integrates the collaborative change information of all nodes in the multi-hop neighborhood. Compared with the original node features, it has stronger feature expression ability and anti-interference ability, and can better distinguish between normal fluctuations of signals and gradual synchronous deception interference in urban canyon environment.

[0039] Step 227: Based on all node features updated through multi-layer aggregation, rearrange them according to the time window order and feature dimensions in the original feature set, and stack the depth features of each node into a three-dimensional tensor. The three dimensions of this tensor correspond to the time window index, signal primitive type, and aggregated feature channels, respectively, to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time steps and spatial channels. Specifically, this includes: determining the two core dimensions of the rearrangement, the time window index and the feature dimensions, ensuring that the rearranged features can retain temporal characteristics and multi-dimensional signal primitive information, and ensuring that the time window index is consistent with the time window order in the original feature set, i.e., the node feature matrix after multi-layer aggregation update. The k-th node feature The first feature corresponds to the k-th time window (k=1 to K) in the original feature set. The arrangement must strictly follow the order of k from 1 to K to ensure that the temporal correlation is not disrupted. The second feature corresponds to the 4L elements of the original combined feature vector. Each element corresponds to a specific positional feature of a signal primitive. The 4L elements are divided into 4 groups, corresponding to four signal primitive types: carrier-to-noise ratio, code phase offset, Doppler frequency shift residual, and signal angle of arrival estimate. Each group contains L elements (corresponding to L sampling points in the sliding time window). The arrangement must strictly follow the element order of the original feature vector to ensure the integrity of the multi-dimensional signal primitive information. All node features after multi-layer aggregation and updating are rearranged to form the node feature matrix. (K×4L) is converted into an intermediate form of a three-dimensional tensor; the specific arrangement process is as follows: Each row (feature vector of each node) The feature is divided into four groups based on its feature dimensions. Each group contains L elements, corresponding to the carrier-to-noise ratio feature group, the code phase offset feature group, the Doppler frequency shift residual feature group, and the signal angle of arrival estimation feature group, denoted as . (Carrier-to-noise ratio, L-dimensional) (Code phase offset, L-dimensional) (Doppler frequency shift residual, L-dimensional) (Signal angle of arrival estimate, L-dimensional), integrate the feature groups of the same signal primitive type across all time windows to obtain four K×L matrices, which are the carrier-to-noise ratio feature matrices. (The element in the k-th row and l-th column is) ), code phase offset feature matrix Doppler frequency shift residual characteristic matrix , Signal angle of arrival estimation characteristic matrix Where l = 1 to L; the depth features of each node are stacked into a three-dimensional composite feature tensor. The three dimensions of the three-dimensional tensor correspond to the time window index, signal primitive type, and aggregated feature channels, respectively. The specific dimensions are defined as follows: the first dimension is the time window index, with a dimension size of K, corresponding to K time windows, in the same order as the original feature set; the second dimension is the signal primitive type, with a dimension size of 4, corresponding to four signal primitives: carrier-to-noise ratio, code phase offset, Doppler frequency shift residual, and signal angle of arrival estimate; the third dimension is the aggregated feature channels, with a dimension size of L, corresponding to L sampling points of the sliding time window, that is, L feature values ​​of each signal primitive in each time window; the stacking process is to stack the 4 K×L feature matrices ( Stacking them along the second dimension in order of signal primitive types yields a three-dimensional composite feature tensor. ,in The eigenvalues ​​(k=1~K, s=1 to 4, l=1 to L) represent the k-th time window, the s-th signal element type, and the l-th feature channel. s=1 corresponds to the carrier-to-noise ratio, s=2 corresponds to the code phase offset, s=3 corresponds to the Doppler frequency shift residual, and s=4 corresponds to the estimated signal angle of arrival. This three-dimensional composite feature tensor fully integrates the cooperative change information of each signal element in the time step (first dimension) and spatial channels (second and third dimensions), and can capture the subtle changes in signal characteristics caused by progressive synchronous deception interference.

[0040] The constructed fully connected feature association graph realizes the structured integration of scattered features; the multi-layer multi-head attention mechanism mines the collaborative change information of each signal primitive in time step and spatial channel through independent learning of multiple attention heads and fusion of multi-hop neighborhood features. The constructed composite feature tensor effectively preserves the temporal gradual change characteristics and multi-dimensional collaborative information of the signal, improving the accuracy and timeliness of deception interference detection.

[0041] In a preferred embodiment of the present invention, step 3 above involves inputting the composite feature tensor into a pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork; simulating a continuous, gradually shifting offset sequence of progressive synchronous deception interference through the generator subnetwork to obtain a pseudo-deception feature tensor; extracting discriminative depth features through forward adversarial discrimination using the discriminator subnetwork based on the pseudo-deception feature tensor and the original composite feature tensor to obtain discriminative depth features; and calculating a deception probability score sequence for the current navigation signal based on the discriminative depth features, which may include: In this embodiment of the invention, step 330 involves inputting the composite feature tensor into the generator sub-network. The generator sub-network uses a one-dimensional temporal convolutional layer to perform a sliding convolution operation on the composite feature tensor in the time dimension, extracting the dynamic change patterns between adjacent frames within each time window, and obtaining an intermediate feature map including temporal dependencies. Specifically, this includes determining the core parameters of the one-dimensional temporal convolutional layer in the generator sub-network. The parameter selection needs to consider the temporal characteristics of the composite feature tensor and the gradual change pattern of the progressive synchronous deception interference to ensure that the dynamic change patterns in the time dimension can be effectively captured. The core parameters include the convolution kernel size. Convolution stride Number of output channels and padding method, where the kernel size Choosing a stride of 3 to 5 (corresponding to 3 to 5 adjacent time windows) can effectively capture the dynamic changes between adjacent time windows; convolution stride Setting it to 1 ensures that feature information in the time dimension is not lost, enabling continuous sliding convolution for each time window; output channel number The feature channel number is selected as 2L to 4L, which is greater than the original composite feature tensor's number of channels L, to expand the feature representation capability and capture richer dynamic change patterns; zero padding is used, with a padding length of [value missing]. To ensure the number of time windows K remains constant after the convolution operation, consistent with the time dimension of the original composite feature tensor, a one-dimensional temporal convolutional layer is constructed. This layer performs sliding convolution operations only on the time dimension (the first dimension of the composite feature tensor), while the signal primitive type dimension and feature channel dimension remain unchanged. This ensures the extraction of dynamic change patterns in the time dimension. The calculation formula for the one-dimensional temporal convolutional layer is as follows: ,in This represents the feature value of the t-th time window, the s-th signal element type, and the l-th feature channel in the intermediate feature map; The kernel size; This represents the weight parameter of the p-th position and the l-th feature channel in the convolution kernel, where the kernel dimension is... (Time dimension × Signal primitive dimension × Feature channel dimension) ensures that convolution is performed only in the time dimension; The first characteristic tensor in the composite feature tensor Given a time window, the s-th signal element type, and the eigenvalue of the l-th feature channel, when... or When zero-filling is used, supplementation is performed. The bias parameter of the l-th feature channel is represented by L; the summation operation is performed along the time dimension of the convolution kernel to achieve sliding convolution; the sliding convolution operation is performed and the intermediate feature map is output, thus combining the composite feature tensor. Given the constructed one-dimensional temporal convolutional layer, sliding convolution is performed sequentially along the time dimension from the first time window to the Kth time window, according to the above calculation formula. Each time window is connected to the adjacent... The features of each time window are convolved to capture the dynamic change patterns between adjacent frames. After the convolution operation is completed, an intermediate feature map is obtained. The feature map retains the original composite feature tensor's time window number K and signal primitive type number 4, while the number of feature channels becomes... It also includes time-domain dependencies in the time dimension, fully reflecting the dynamic changes of each signal element between adjacent time windows.

[0042] Step 331: Input the intermediate feature map into the channel compression convolutional layer. The channel compression convolution reduces the number of channels in the intermediate feature map to a preset low-dimensional latent space, thus obtaining the conditional latent code. Specifically, this includes: setting the preset low-dimensional latent space dimension. The choice of dimension needs to balance feature compression efficiency and key feature preservation, and is usually set to L to 2L. This can effectively reduce the computational complexity of subsequent deconvolutional networks while retaining key information reflecting the dynamic changes of the signal in the intermediate feature maps, avoiding information loss caused by feature compression. For example, when the original number of feature channels L=20, A value of 30 can be selected to achieve reasonable dimensionality compression. A channel compression convolutional layer is constructed. The core function of this layer is to compress the feature channels of the intermediate feature map without changing the number of time windows or signal primitive types; it only reduces the feature channel dimension. The parameters of the channel compression convolutional layer are set as follows: the kernel size is set to 1×1×1 (time dimension × signal primitive dimension × feature channel dimension) to ensure that only the feature channels are compressed, without affecting the feature information in the time dimension and signal primitive dimension; the convolution stride is set to 1 to avoid feature sampling loss; and the number of output channels is set to the preset low-dimensional latent space dimension. The padding method uses zero padding to ensure that the number of time windows K and the number of signal primitive types 4 remain unchanged; the calculation formula for the channel-compressed convolutional layer is as follows: ,in This represents the feature value of the t-th time window, the s-th signal primitive type, and the d-th latent space dimension in the conditional implicit coding. The number of output channels for the intermediate feature map; The weight parameters of the channel-compressed convolutional layer are represented by the dimension . , used to 3D feature channel mapping to Hidden space; This represents the feature value of the t-th time window, the s-th signal element type, and the l-th feature channel in the intermediate feature map; This represents the bias parameter for the d-th latent space dimension, where the dimension is... The summation operation is performed along the feature channel dimensions of the intermediate feature map, achieving channel compression; then, a channel compression convolution operation is performed to compress the intermediate feature map. The pre-constructed channel-compressed convolutional layer is input, and according to the above calculation formula, the feature channels of each time window and each signal primitive type are compressed and mapped. 3D feature channel mapping to After compression, the conditional implicit code is obtained from the low-dimensional latent space. This hidden encoding preserves the temporal dependencies and signal primitive features in the intermediate feature map, while significantly reducing the feature dimension and removing redundant information.

[0043] Step 332 involves inputting the conditional implicit code into a multi-layer deconvolutional network for progressive upsampling and reconstruction. Each deconvolutional layer is followed by a batch normalization layer and a non-linear activation function to gradually restore the same time window length and number of feature channels as the original composite feature tensor, resulting in a preliminary reconstructed tensor. Specifically, this includes constructing a multi-layer deconvolutional network with a specified number of layers. Based on the dimensions of conditional implicit coding The difference in the number of feature channels L from the original composite feature tensor is determined, typically taking a value of 2 to 4 layers to ensure gradual upsampling and smooth recovery to the target number of feature channels, avoiding feature distortion during upsampling. Each deconvolutional network layer consists of a deconvolution layer + batch normalization layer + nonlinear activation function. The parameters of each layer are set independently, gradually increasing the number of feature channels until it matches the number of feature channels L of the original composite feature tensor. The number of time windows K and the number of signal primitive types 4 remain constant. The parameter settings of each deconvolutional layer follow the principle of gradual upsampling. The number of input channels in the first deconvolutional layer is... The number of output channels is set to ×2, the kernel size is set to 3×1×1 (time dimension × signal primitive dimension × feature channel dimension), the stride is set to 1, and the padding method is zero padding to ensure that the number of time windows and the number of signal primitive types remain unchanged; the number of input channels of each subsequent deconvolution layer is equal to the number of output channels of the previous layer, and the number of output channels gradually increases until the number of output channels of the last deconvolution layer equals L, thus achieving complete recovery of the number of feature channels. For example, when When L=30 and L=20, if a two-layer deconvolutional network is set, the number of output channels of the first layer is set to 40 and the number of output channels of the second layer is set to 20, and the channel number is restored.

[0044] Perform a step-by-step upsampling reconstruction operation to implicitly encode the conditions. The input is the first deconvolution layer, and the formula for calculating the deconvolution layer is: ,in This represents the feature value of the t-th time window, the s-th signal primitive type, and the l-th feature channel in the output features of the first deconvolution layer; This represents the weight parameters of the first deconvolutional layer, with dimension 1. ( (Number of output channels in the first layer). This represents the feature value of the t-th time window, the s-th signal primitive type, and the d-th latent space dimension in the conditional implicit coding. This represents the bias parameters of the first deconvolutional layer, with dimension 1. The summation operation is performed along the latent space dimension, achieving upsampling and increasing the number of channels. After the deconvolution operation, the output features are input into the batch normalization layer. The purpose of the batch normalization layer is to standardize the feature distribution, accelerate network training, and avoid gradient vanishing. The calculation formula is as follows: ,in Represents the eigenvalues ​​after batch normalization. This represents the mean of the l-th feature channel. This represents the variance of the l-th feature channel. For a small constant (with a value of...) To avoid a denominator of 0, after batch normalization, a non-linear activation function is input, using the LeakyReLU function, calculated as follows: It can preserve the weak features of negative values, improve feature representation ability, and avoid gradient vanishing.

[0045] The output features of the first deconvolutional layer are input into the second deconvolutional layer. The process of deconvolution + batch normalization + Leaky ReLU activation is repeated, gradually increasing the number of feature channels until all features are activated. The upsampling reconstruction of the deconvolutional network, the features output by the last deconvolutional network, has a time window size of [number missing]. The signal primitive types are 4, and the number of feature channels is L, which is completely consistent with the dimension of the original composite feature tensor. This feature is used as the initial reconstructed tensor, denoted as . The tensor has recovered the dimensions of the original composite feature tensor, but has not yet been subjected to offset modulation with deceptive interference.

[0046] Step 333: At each time step of the initially reconstructed tensor, an offset modulation is applied to the output feature vector, which increases linearly with the window index. This ensures that the feature vector corresponding to each time window continuously and smoothly evolves from the normal baseline value of the starting window to the offset target value of the ending window. This simulates the slow drift process of the carrier-to-noise ratio, code phase, Doppler frequency shift, and signal angle of arrival parameters caused by progressive synchronous deception interference, ultimately obtaining a pseudo-deception feature tensor with dimensions completely consistent with the original composite feature tensor. Specifically, this includes: determining the core parameters of the offset modulation, including the normal baseline value, the offset target value, and the offset coefficient. The parameter settings must conform to the characteristics of progressive synchronous deception interference, simulating the slow drift of the carrier-to-noise ratio, code phase offset, Doppler frequency shift residual, and signal angle of arrival estimate caused by interference, ensuring that the pseudo-deception feature tensor can realistically simulate the continuous and gradual process of deception interference. The normal baseline value is the initially reconstructed tensor. The eigenvector of the starting time window (t=1) is denoted as... (s=1 to 4, l=1 to L), corresponding to the normal signal characteristics without deception interference; the offset target value is set according to the typical offset range of progressive synchronous deception interference, and different offset targets are set for the four signal primitives, denoted as... (s=1 to 4, l=1 to L), where the target offset value for the carrier-to-noise ratio is 5 to 10 dB-Hz lower than the normal baseline value (corresponding to an offset of 0.1 to 0.2 after normalization), the target offset value for the code phase offset is 0.1 to 0.2 chips lower than the normal baseline value (corresponding to an offset of 0.2 to 0.4 after normalization), the target offset value for the Doppler frequency shift residual is 5 to 8 Hz lower than the normal baseline value (corresponding to an offset of 0.25 to 0.4 after normalization), and the target offset value for the estimated angle of arrival is lower than the normal baseline value. to (Normalized offsets range from 0.1 to 0.15); Offset coefficient The time window index increases linearly with time window t, and the calculation formula is as follows: , where t is the time window index (t=1 to K). Number of time windows The range of values ​​for is [0, 1], and when t=1 =0 (no offset), at t=K =1 (maximum offset), achieving linear increment of the offset.

[0047] Perform offset modulation operation on the initially reconstructed tensor At each time step, the output feature vector is subjected to offset modulation, and the modulation calculation formula is as follows: ,in The eigenvalues ​​of the t-th time window, the s-th signal primitive type, and the l-th feature channel in the spoofing feature tensor are represented. This represents the eigenvalue at the corresponding position in the initially reconstructed tensor; This represents the offset coefficient for the t-th time window; This represents the offset target value of the s-th signal element and the l-th feature channel; This represents the normal baseline value of the s-th signal element and the l-th characteristic channel; - () represents the maximum offset, and the offset coefficient. Multiplication yields the actual offset of the current time window. During offset modulation, the feature vector corresponding to each time window starts from the normal baseline value (no offset) of the initial window (t=1). As the time window index t increases, the offset gradually increases until the target offset value is reached in the final window (t=K). This achieves a continuous and smooth evolution of the feature vector, simulating the slow drift process of carrier-to-noise ratio, code phase offset, Doppler frequency shift residual, and signal angle of arrival estimate caused by progressive synchronous deception interference, thus avoiding abrupt offset changes (consistent with the characteristics of progressive deception interference). After offset modulation of all time windows is completed, the pseudo-deception feature tensor is obtained. The number of time windows, the number of signal primitive types, and the number of feature channels of this tensor are completely consistent with the original composite feature tensor. Only the linearly increasing offset of the simulated deception interference is applied to the eigenvalues, which can truly reflect the feature change law of the gradual synchronous deception interference.

[0048] Step 334: Based on the pseudo-spoofing feature tensor and the original composite feature tensor, the discriminator sub-network concatenates the features of the two along the channel dimension to obtain the combined feature tensor; specifically, this includes: determining the dimensionality consistency of the two input tensors, and the original composite feature tensor... False deception feature tensor The number of time windows for both The number of signal primitive types (4) and the number of feature channels (L) are completely consistent to ensure effective splicing in the channel dimension and prevent dimensional misalignment after splicing. The splicing dimension and order are determined: the splicing dimension is the feature channel dimension (the third dimension of the composite feature tensor), and the splicing order is the original composite feature tensor first, followed by the pseudo-spoofing feature tensor. This ensures the discriminator sub-network can clearly distinguish between real and pseudo-spoofing features, facilitating subsequent extraction of their differences and improving discrimination accuracy. During splicing, the time window dimension and signal primitive type dimension remain unchanged; only the feature channels of the two tensors are merged, resulting in 2L feature channels after splicing (original feature channel number L + pseudo-spoofing feature channel number L). The specific splicing operation is as follows: for each time window t (t=1 to K) and each signal primitive type s (s=1 to 4), the L-dimensional feature channels of the original composite feature tensor for that time window and signal primitive type are... The L-dimensional feature channels corresponding to the time window and signal primitive type in the pseudo-deception feature tensor. By splicing, a 2L-dimensional feature channel is obtained. After all time windows and signal primitive types of feature channels are spliced ​​together, a combined feature tensor is obtained. Where K is the number of time windows, 4 is the number of signal primitive types, and 2L is the number of feature channels after splicing. This combined feature tensor contains both the composite features of the original normal signal and the pseudo-spoofing features of the simulated spoofing interference. The difference between the two features can provide a clear basis for discrimination for the discriminator sub-network, making it easier to extract discriminative deep features and distinguish between normal signals and spoofing interference signals.

[0049] Step 335: Based on the combined feature tensor, the discriminator sub-network calculates attention weights along the time dimension and feature channel dimension using multiple parallel spatiotemporal attention units, and weights and aggregates the features in each dimension according to the attention weights to obtain preliminary deep features. The preliminary deep features obtained from each spatiotemporal attention unit are then fused element-wise to obtain the fused deep features. Specifically, this includes: constructing multiple parallel spatiotemporal attention units, the number of attention units... Typically, the number of attention weights is 2 to 4. Each attention unit independently learns attention weights in different dimensions, capturing key features in the time dimension and feature channel dimension respectively. The comprehensiveness of feature aggregation is improved through multi-unit fusion. Each spatiotemporal attention unit contains a time attention module and a channel attention module. The two modules are computed in parallel, outputting attention weights in the time dimension and channel dimension respectively. The two weights are then combined for feature weighted aggregation. The time attention module is used to calculate attention weights along the time dimension, highlighting key dynamic changes between time windows, especially the temporal gradual changes caused by deception interference. The calculation process is as follows: first, the combined feature tensor is processed... Global average pooling is performed along the signal primitive type and feature channel dimensions to obtain the feature vector in the time dimension. The calculation formula is: ,in Let represent the time feature value of the t-th time window. The summation operation covers all signal primitive types and feature channels to obtain the global feature of each time window. Subsequently, Input a two-layer fully connected network. The first fully connected layer maps K-dimensional temporal features to K / 2 dimensions using ReLU activation. The second fully connected layer maps back to K dimensions using Sigmoid activation, yielding the temporal attention weights. The value range is [0, 1]. The larger the weight, the more critical the feature corresponding to the time window. The channel attention module is used to calculate the attention weight along the feature channel dimension, highlighting the key channel features that can distinguish normal signals from deceptive interference. The calculation process is as follows: first, the combined feature tensor is processed... Global average pooling is performed along the time window and signal primitive type dimensions to obtain the feature vector along the channel dimension. The calculation formula is: ,in Let l represent the channel eigenvalue of the l-th characteristic channel. The summation operation covers all time windows and signal primitive types, yielding the global feature of each characteristic channel. Input a two-layer fully connected network. The first fully connected layer maps the 2L-dimensional channel features to L-dimensional features using ReLU activation. The second fully connected layer maps back to 2L-dimensional features using Sigmoid activation, yielding the channel attention weights. The value ranges from [0, 1]. The larger the weight, the stronger the discrimination ability of the corresponding feature channel.

[0050] Single attention unit feature aggregation, including temporal attention weights and channel attention weights With combined feature tensor Element-wise multiplication is performed to achieve weighted feature aggregation, resulting in the preliminary deep features of a single spatiotemporal attention unit. The calculation formula is as follows: ,in For the initial deep features of a single attention unit, a weighted operation is used to highlight the feature information of key time windows and key feature channels while suppressing redundant information; feature fusion of multiple attention units... The initial depth features obtained from each parallel spatiotemporal attention unit are fused element-wise by addition. The fusion calculation formula is as follows: ,in For the fused deep features, The initial deep features of the nth attention unit are added and fused element by element, which can integrate the key features of different dimensions captured by multiple attention units, thereby improving the expressive and discriminative power of deep features.

[0051] Step 336: Based on the fused depth features, the discriminator subnetwork performs a global average pooling operation along the feature channel dimension to compress channel information while retaining the complete length of the time dimension, resulting in discriminative depth features, which are two-dimensional tensors. Specifically, this includes: determining the operation dimension and purpose of global average pooling; the operation dimension is the feature channel dimension (the third dimension of the fused depth features); the purpose is to compress redundant information in the feature channel dimension, retaining the core features of each time window and each signal primitive type, and simultaneously converting the three-dimensional fused depth features into two-dimensional discriminative depth features for easier processing by the fully connected layer and calculation of the deception probability score. During global average pooling, the time window dimension and signal primitive type dimension remain unchanged; only the feature channel dimension is pooled and compressed. The global average pooling operation is then performed, and the calculation formula for global average pooling is... ,in This represents the feature value of the t-th time window and the s-th signal primitive type in the discriminative deep feature; This represents the number of feature channels for the fused deep features. Let represent the feature value of the t-th time window, the s-th signal primitive type, and the l-th feature channel in the fused deep features. The summation operation is performed along the feature channel dimension, and then the average value is taken to obtain the core feature value for each time window and each signal primitive type, thus compressing the channel information. After the pooling operation, the discriminative deep features are obtained. The feature is a two-dimensional tensor. The first dimension is the time window index (size K), and the second dimension is the signal primitive type (size 4). Compared with the fused deep feature, the dimensions are greatly compressed, removing redundant information from the feature channels, while retaining the full length of the time dimension and the core discriminative features of the four signal primitives. It can reflect the key changes of each signal primitive within each time window.

[0052] Step 337: Based on the discriminative deep features, the discriminator sub-network inputs the feature vector corresponding to each time step in the discriminative deep features into a fully connected layer with shared weights. This fully connected layer has a single output node and undergoes a time-step nonlinear mapping using a Sigmoid activation function to obtain a series of deception probability scores between 0 and 1. The scores of all time steps are arranged in chronological order to form a deception probability score sequence. Specifically, this includes constructing a fully connected layer with shared weights. The core function of this fully connected layer is to map the feature vector corresponding to each time step in the discriminative deep features to a single deception probability score, and all time steps share the same set of weight parameters to ensure the consistency of the scoring standard and avoid excessive deviation in scores between different time steps. The parameters of the fully connected layer are set as follows: input dimension is 4 (corresponding to four signal primitive types), output dimension is 1 (single output node), and weight matrix... Bias vector The weights and bias parameters are learned synchronously during the pre-training of the adversarial neural network to ensure that the relationship between discriminative deep features and deception probability can be mapped; time-step feature mapping is performed for discriminative deep features. Each time window t (t=1 to K) corresponds to a 4-dimensional feature vector. (Core feature values ​​of four signal primitive types), input this feature vector into the shared weighted fully connected layer, and the calculation formula is as follows: ,in This represents the linear mapping result for the t-th time window, and is a scalar value. This is the weight matrix of the fully connected layer; It is the transpose (column vector) of the 4-dimensional feature vector of the t-th time window. The bias vector of the fully connected layer has no fixed range of values ​​for its linear mapping result, which needs to be mapped to the [0, 1] interval by a subsequent activation function. The linear mapping result is then input into the Sigmoid activation function for non-linear mapping, converting the result into a deception probability score between 0 and 1. The formula for calculating the Sigmoid activation function is as follows: ,in This represents the deception probability score for the t-th time window, with a value ranging from [0, 1]. The value is a natural constant (approximately 2.71828). A score closer to 1 indicates a higher likelihood of the navigation signal being subject to progressive synchronization-type deception interference within that time window; a score closer to 0 indicates a signal closer to normal operation and no deception interference. The deception probability scores for all time windows are arranged chronologically, i.e., according to the time window index t from 1 to K. Arranged sequentially, forming a deception probability scoring sequence. This sequence fully reflects the probability of deception interference of navigation signals in each time window, and can clearly present the gradual development process of deception interference (from low probability to high probability), providing a direct basis for judging whether deception interference exists, determining the interference start time and interference intensity based on the scoring sequence.

[0053] The generator subnetwork simulates the continuous and gradual process of deception interference through temporal convolution, channel compression, and deconvolution reconstruction, combined with linearly increasing offset modulation. The generated pseudo-deception feature tensor is highly consistent with the real features. The discriminator subnetwork reduces false detections, false negatives, and detection delays through feature concatenation, spatiotemporal attention aggregation, and global pooling.

[0054] In a preferred embodiment of the present invention, step 4 above, comparing the deception probability score sequence with a preset detection threshold, and determining that a progressive synchronous deception interference exists if the deception probability score is higher than the detection threshold, and generating an alarm signal, may include: In this embodiment of the invention, step 440 involves smoothing the deception probability score sequence calculated over multiple consecutive time windows using a moving average filter to obtain a smoothed score value sequence. Specifically, this includes determining the input deception probability score sequence, which is the output of step 337. ,in The deception probability score for the t-th time window is given, with a value range of [0, 1], where K is the number of time windows. The score sequence may contain abnormally high or low values ​​due to instantaneous signal fluctuations, requiring smoothing to eliminate such interference. A moving average filter is constructed, and the core parameters of the filter and the size of the moving window are determined. The selection of the sliding window size needs to consider both the number of time windows K and the asymptotic characteristics of the deception interference. Typically, a value of 3 to 5 time windows is chosen. This effectively filters out random noise without over-smoothing and losing the asymptotic trend of the deception interference, ensuring that the smoothed sequence truly reflects the changing pattern of the deception probability. The moving average filter uses a simple moving average method, where each smoothed score is calculated from the current time window and the adjacent previous values. The deception probability scores for each time window are averaged. A moving average smoothing operation is performed. Specifically, starting from the first time window, when the time window index t is less than the moving window size w, the average of the scores from the previous t time windows is used as the smoothed score value for the current time window. That is, when t=1, the smoothed score value is... When t=2 and w=3, the smoothed score value When the time window index t is greater than or equal to the sliding window size w, the current time window and the previous time window are used. The average score of each time window is used as the smoothed score for the current time window. The specific calculation method is to compare the deception probability score of the current time window with the previous time window. The deception probability scores for each time window are summed sequentially to obtain the total score. This total is then divided by the sliding window size w to obtain the smoothed score value for the current time window. This process is repeated for each time window in the deception probability score sequence until all K time windows have been smoothed, resulting in the final smoothed score value sequence. The number of time windows for this sequence remains K, and each smoothing score value The value range of is still [0, 1], and the fluctuation range has been greatly reduced, which can reflect the gradual change trend of the probability of deception.

[0055] Step 441: Compare each score value in the smoothed score value sequence with a preset detection threshold. If the number of scores in the smoothed score value sequence that continuously exceed the detection threshold reaches a preset confirmation threshold, it is determined that there is progressive synchronous deception interference in the current navigation signal. Specifically, this includes: setting a preset detection threshold and a confirmation threshold. The detection threshold is used to distinguish between normal signals and deception interference signals. It is set based on a large amount of experimental data and the requirements of UAV navigation scenarios, with a value range of [0.5, 0.7]. It is usually set to 0.6, that is, when the smoothed score value is greater than 0.6, it is initially determined that there may be deception interference in this time window. The confirmation threshold is used to avoid misjudgment caused by instantaneous fluctuations. It is combined with the sliding window size and the progressive nature of the deception interference. The input characteristic setting is typically set to 3 to 5 time windows. This means that deceptive interference is only confirmed when the smoothed score values ​​for 3 to 5 consecutive time windows exceed the detection threshold, ensuring the reliability of the judgment result. A spherical kinematic attitude calculation algorithm is constructed to calculate the real-time attitude information of the UAV, assisting in verifying the interference judgment result and avoiding misjudgments caused by fluctuations in the navigation signal itself. The spherical kinematic attitude calculation algorithm takes the angular velocity and acceleration data collected by the UAV's built-in inertial measurement unit as input. Based on the attitude motion equations in the spherical coordinate system, it calculates the UAV's pitch angle, roll angle, and yaw angle—three attitude parameters. The specific construction and implementation process is as follows: Input data acquisition: Obtain real-time data collected by the UAV's inertial measurement unit, including three-axis angular velocities. (Corresponding to the angular velocities of the UAV along the x, y, and z axes respectively) and three-axis acceleration (Corresponding to the acceleration along the x, y, and z axes of the UAV respectively), the acquisition frequency is kept consistent with the time window frequency, that is, each time window corresponds to a set of inertial measurement data, ensuring that the attitude calculation and the deception probability score are synchronized in time; initial attitude initialization, setting the initial attitude parameters of the UAV, initial pitch angle Initial roll angle Initial heading angle Assuming the UAV's initial state is horizontal, uniform flight, and stable attitude, this provides an initial reference for attitude iteration calculation. Attitude increments are calculated by integrating angular velocity. Based on the spherical kinematics attitude equations, the three-axis angular velocities are integrated over time to calculate the attitude increment within each time window. Specifically, the pitch angle increment is calculated as follows: Roll angle increment Heading angle increment ,in The time interval of the time window, i.e., the time difference between two adjacent time windows, is used to multiply the angular velocity of each axis by the time interval to obtain the corresponding attitude increment within that time window. Attitude iterative update involves iteratively updating the UAV's real-time attitude parameters based on the initial attitude parameters and the attitude increments for each time window. Specifically, the pitch angle of the current time window is used as the basis for the update. The roll angle of the current time window The heading angle of the current time window That is, the attitude parameters of the previous time window are added to the attitude increment of the current time window to obtain the real-time attitude parameters of the current time window. This process is repeated to obtain the real-time attitude parameters of the UAV for each time window.

[0056] Attitude verification and correction involves comparing the calculated real-time attitude parameters with the attitude thresholds during normal UAV flight. During normal flight, the pitch and roll angles typically range from [value missing]. The range of values ​​for the heading angle is: If the attitude parameters of a certain time window exceed the threshold range, it indicates that the UAV's flight attitude is abnormal, possibly due to navigation deviation caused by deception interference. In this case, the weight of the deception interference judgment for that time window is increased. If the attitude parameters are within the normal threshold range, it indicates that the UAV's flight attitude is stable, which helps to confirm the reliability of the deception probability score for that time window and avoid misjudgment. The deception interference judgment operation is then performed. The specific process is as follows: first, the smoothed score value sequence is... Each smoothing score value The results are compared one by one with the preset detection thresholds, and the comparison results for each time window are recorded. If the number of consecutive suspected interference windows exceeds the detection threshold, the time window is marked as a suspected interference window; otherwise, it is marked as a normal window. Next, the number of consecutive suspected interference windows is counted, starting from the first suspected interference window. If the number of consecutive suspected interference windows reaches a preset confirmation threshold, the real-time attitude information obtained from the spherical kinematics attitude calculation algorithm is used for verification. If the UAV attitude parameters exceed the normal threshold range, or the rate of change of the attitude parameters exceeds the normal range (i.e., the absolute value of the attitude increment in adjacent time windows is greater than the preset increment threshold), then deceptive interference is further confirmed. If the attitude parameters are normal, the number of consecutive suspected interference windows is checked again to ensure no misjudgment. If, after the above comparison and verification, the number of consecutive scores exceeding the detection threshold in the smoothed score value sequence reaches the preset confirmation threshold, and the attitude calculation results are verified to be correct, then it is determined that there is progressive synchronous deceptive interference in the current navigation signal, completing the interference determination.

[0057] Step 442: In response to the determination of the existence of progressive synchronous deception interference, an alarm signal including interference type identifier and interference start time information is obtained. Specifically, this includes: determining the core components of the alarm signal, including interference type identifier and interference start time information. Both types of information are extracted based on the determination result and related data from step 441, ensuring the accuracy and practicality of the information, without adding irrelevant information, and meeting the actual needs of interference detection; extracting the interference type identifier, which is used to clarify the currently detected interference type. Since it targets progressive synchronous deception interference, the interference type identifier is set to a fixed identifier, specifically progressive synchronous deception interference. This identifier is consistent with the corresponding identifier in the preset interference type library, facilitating the UAV navigation system to identify the interference type and take targeted anti-interference measures; determining the interference start time information. The interference start time is the actual time corresponding to the first consecutive time window exceeding the detection threshold in the smoothed score value sequence. The specific extraction process is as follows: first, find the start time window index of the consecutive suspected interference windows counted in step 441, denoted as... That is, the first one to satisfy The time window index is defined as the time window that exceeds the detection threshold and where multiple consecutive time windows subsequently meet this condition. Based on the correspondence between the time window index and the actual time, the actual time corresponding to that time window, i.e., the interference start time, is calculated. Specifically, the interference start time is... ,in This refers to the actual start time corresponding to the first time window. To disrupt the start time window index, The time interval of the time window is calculated by adding the product of (interference start time window index minus 1) and the time interval to the actual start time of the first time window, thus obtaining the interference start time and locating the moment when the interference begins. The interference type identifier and interference start time information are integrated to generate a complete alarm signal. The alarm signal adopts a preset signal format to ensure that the UAV navigation system can quickly identify and analyze it. The signal clearly indicates that the interference type is progressive synchronous deception interference, and also indicates the specific interference start time, so that operators can understand the interference situation in a timely manner and take emergency anti-interference measures, such as switching navigation modes and adjusting flight attitude, to minimize the impact of deception interference on UAV navigation.

[0058] By smoothing out random fluctuations in the deception probability score, integrating spherical kinematics attitude calculation algorithms to assist in verifying interference judgment results, and combining detection thresholds and confirmation count thresholds to determine deception interference and generate alarm signals, the accuracy of interference judgment is improved, avoiding false and false judgments; effectively reducing the threat of progressive synchronous deception interference to UAV navigation safety.

[0059] like Figure 2As shown, embodiments of the present invention also provide a UAV satellite navigation spoofing and interference detection system based on adversarial neural networks, comprising: The feature extraction module is used to collect satellite navigation baseband signals in real time in the multipath reflection micro-environment of urban canyons where the ratio of building height to street width is greater than 2, for satellite navigation receiver terminals mounted on the belly of UAVs with the antenna phase center located above the fuselage structure. The module extracts multi-dimensional signal primitive features of the satellite navigation baseband signals, performs zero-mean unit variance normalization and nonlinear dimensionality reduction mapping on each dimension of the signal primitive features to obtain the corresponding feature vectors, and aligns each feature vector along the time axis by sliding window and splices it with multi-dimensional tensors to form the original feature set. The fusion module is used to calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology. A multi-head attention mechanism is used to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time step and spatial channel. The deception detection module is used to input a composite feature tensor into a pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork. The generator subnetwork simulates a continuous, gradually shifting sequence of progressive synchronous deception interference to obtain a pseudo-deception feature tensor. Based on the pseudo-deception feature tensor and the original composite feature tensor, discriminative deep features are extracted through forward adversarial discrimination by the discriminator subnetwork. Based on the discriminative deep features, a deception probability score sequence for the current navigation signal is calculated. The alarm generation module compares the deception probability score sequence with a preset detection threshold. If the deception probability score is higher than the detection threshold, it determines that there is progressive synchronous deception interference and generates an alarm signal.

[0060] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0061] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0062] Experimental Example: Progressive Synchronous Deception Interference Detection Experiment for UAVs in an Urban Canyon Environment Experimental scenario: The drone performs a logistics delivery mission in the canyon area of ​​XX city (39.91°N, 116.40°E), with a flight altitude of 50m and a speed of 5m / s. It is equipped with a Beidou B1C band receiver (sampling frequency 10Hz) and the experiment lasts for 120s. From the 60th second, a progressive synchronous deception interference is injected (the false signal power increases by 0.5dB every 10s, the code phase offset rate is 0.01chip / s, and the Doppler frequency shift residual drift is 0.1Hz / s).

[0063] Step 1: Extraction of multi-dimensional signal primitive features and construction of the original feature set Experimental data: Four-dimensional features were extracted: carrier-to-noise ratio, code phase offset, Doppler frequency shift residual, and signal angle of arrival, totaling 1200 sampling points (120s×10Hz).

[0064] Sliding window parameters: window length L=20 (2s), sliding step S=10 (1s), generating K=119 time windows.

[0065] Key results:

[0066] Step 2: Construction of Feature Association Graph and Generation of Composite Feature Tensor Experimental data: Mutual information calculation, number of bins B=5, carrier-to-noise ratio - code phase mutual information I=0.85 (high correlation), signal angle of arrival - Doppler residual I=0.62 (medium correlation).

[0067] Temporal dependency strength: maximum time delay =20, code phase-Doppler residual timing dependence D=0.78 (strong coordination).

[0068] Standardized association weights: After sigmoid compression, the weight range The average weight is 0.56.

[0069] Key results: The fully connected feature graph contains 119 nodes and a total of 7021 edges, with the edge weights distributed in a right-skewed manner (70% of the edges have a weight > 0.4).

[0070] Composite feature tensor dimensions: 119×4×20 (time window×signal primitive×feature channel).

[0071] Step 3: Adversarial neural network detection and deception probability scoring Experimental data: Generator Subnetwork: One-Dimensional Convolutional Kernel =3, Output Channels =40, low-dimensional hidden space =30; Offset modulation coefficient It increases linearly from 0 to 1 (60-120s).

[0072] Discriminator subnetwork: number of spatiotemporal attention units =3, Deception Probability Score = .

[0073] Key results: No-deception phase (0-60s): Average score =0.12±0.05; Deception phase (60-120s): The score rises from 0.25 to 0.88, and the detection threshold of 0.6 is exceeded at 85s.

[0074] like Figure 3 The graph shown is a deception probability score curve (horizontal axis: time / s, vertical axis: score, red dashed line is the threshold of 0.6, marking the process of the score increasing from a gradual increase to a sudden jump).

[0075] Step 4, Alarm Generation and Verification Experimental data: Moving average filter window Smoothed score The volatility decreased by 60%; Confirmation threshold = 3, 3 consecutive windows starting from the 87th second. >0.6, trigger an alarm; Spherical kinematic attitude calculation: Pitch angle increment during the deception phase (Normal <1° / s), auxiliary verification shows interference exists.

[0076] Key results: Alarm delay: 27 seconds from the start of the deception (60s) to the alarm trigger (87s), earlier than traditional carrier-to-noise ratio single feature detection (45s); False positive rate: 2 false positives in 100 interference-free experiments, with a false positive rate of 2%.

[0077] like Figure 4 The graph shows a comparison between the smoothing score and the detection threshold (horizontal axis: time / s, vertical axis: smoothing score, red shaded area represents continuous periods exceeding the threshold).

[0078] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks, characterized in that, The method includes: Step 1: For the satellite navigation receiver terminal mounted on the belly of the UAV with the antenna phase center located above the fuselage structure, in the multipath reflection micro-environment of urban canyons where the ratio of building height to street width is greater than 2, the satellite navigation baseband signal is collected in real time, and the multi-dimensional signal primitive features of the satellite navigation baseband signal are extracted. The signal primitive features of each dimension are normalized with zero mean and unit variance and nonlinear dimensionality reduction mapping are performed to obtain the corresponding feature vectors. The feature vectors are aligned by sliding window along the time axis and spliced ​​with multi-dimensional tensors to form the original feature set. Step 2: Calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology; use a multi-head attention mechanism to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time step and spatial channel. Step 3: Input the composite feature tensor into the pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork; simulate the continuous gradient offset sequence of progressive synchronous deception interference through the generator subnetwork to obtain the pseudo-deception feature tensor; extract discriminative deep features through forward adversarial discrimination of the discriminator subnetwork based on the pseudo-deception feature tensor and the original composite feature tensor to obtain the discriminative deep features; calculate the deception probability score sequence for the current navigation signal based on the discriminative deep features. Step 4: Compare the deception probability score sequence with the preset detection threshold. If the deception probability score is higher than the detection threshold, it is determined that there is a progressive synchronous deception interference and an alarm signal is generated.

2. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 1, characterized in that, The multidimensional signal primitive features include carrier-to-noise ratio time series, code phase offset, Doppler frequency shift residual, and signal angle of arrival estimate.

3. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 1, characterized in that, Calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology, including: Extract any two feature vectors from the original feature set, and discretize and bin the sampling points of these two feature vectors within the same time window. Calculate the joint occurrence frequency and the single-dimensional occurrence frequency in each bin. Calculate the mutual information between the two feature vectors based on the ratio of joint frequency to single-dimensional frequency to obtain the first intermediate parameter reflecting the strength of their statistical correlation. Based on the first intermediate parameter and the value change sequence of the same pair of feature vectors between adjacent time windows, the cross-covariance function of the two feature vector sequences on the time axis is calculated, and the cross-covariance function is normalized to its maximum value within the specified maximum time delay range to obtain the time dependence strength reflecting the degree of mutual following change of the two in continuous time steps, which is used as the second intermediate parameter. The first intermediate parameter is multiplied by the preset first weight coefficient, the second intermediate parameter is multiplied by the preset second weight coefficient, and the two products are added together to obtain the comprehensive association score. The comprehensive association score is then mapped to the preset intensity range using a nonlinear compression transformation to obtain the standardized association weight between each pair of feature vectors. Based on the standardized association weights and all feature vectors in the original feature set, each feature vector is treated as a graph node, and the standardized association weight between each pair of nodes is used as the weight of the edge. Connecting edges are established between all pairs of nodes to construct a feature association graph with a fully connected topology.

4. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 3, characterized in that, A multi-head attention mechanism is employed to perform multi-level aggregation and updating of node features in the feature association graph, resulting in a composite feature tensor that integrates the collaborative change information of each signal primitive across time steps and spatial channels. This tensor includes: The feature vector of each node in the feature association graph is mapped to a query vector, a key vector, and a value vector, respectively. The similarity score between the current node and the query vector and key vector of all other nodes under each attention head is calculated. The similarity score is normalized to obtain the attention weight between the current node and each neighboring node. The value vectors of all nodes are weighted and summed according to the attention weights to obtain the single-head aggregated features of the current node under the attention head. The single-head aggregated features output by all attention heads of each node are concatenated or averaged and then subjected to a nonlinear transformation to obtain the node features updated by the first layer graph attention network. Based on the node features updated by the first-layer graph attention network, the node feature aggregation and update operation of the multi-layer graph attention network is repeatedly performed. The input of each layer is the node features output by the previous layer. Each layer independently learns different attention parameters, so that the final features of each node are integrated with the collaborative change information of all nodes in the multi-hop neighborhood, resulting in all node features after multi-layer aggregation and update. Based on the node features updated through multi-layer aggregation, they are rearranged according to the time window order and feature dimensions in the original feature set. The depth features of each node are stacked into a three-dimensional tensor. The three dimensions of this tensor correspond to the time window index, signal primitive type, and aggregated feature channels, respectively, resulting in a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time steps and spatial channels.

5. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 4, characterized in that, The composite feature tensor is input into a pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork. The generator subnetwork simulates a continuous, gradually shifting sequence of progressively synchronous deception interference to obtain a pseudo-deception feature tensor, including: The composite feature tensor is input into the generator sub-network. The generator sub-network uses a one-dimensional temporal convolutional layer to perform a sliding convolution operation on the composite feature tensor in the time dimension, extracting the dynamic change pattern between adjacent frames within each time window, and obtaining an intermediate feature map including temporal dependencies. The intermediate feature map is input into the channel compression convolutional layer. The channel compression convolution reduces the number of channels in the intermediate feature map to a preset low-dimensional latent space, thus obtaining the conditional latent code. The conditional implicit code is input into a multi-layer deconvolutional network and upsampled and reconstructed step by step. Each deconvolutional operation is followed by a batch normalization layer and a non-linear activation function to gradually restore the same time window length and number of feature channels as the original composite feature tensor, thus obtaining the initial reconstructed tensor. At each time step of the initial reconstruction tensor, an offset modulation that increases linearly with the window index is applied to the output feature vector. This makes the feature vector corresponding to each time window evolve continuously and smoothly from the normal baseline value of the starting window to the offset target value of the ending window. This simulates the slow drift process of carrier-to-noise ratio, code phase, Doppler frequency shift and signal angle of arrival parameters caused by progressive synchronous deception interference. Finally, a pseudo-deception feature tensor with dimensions completely consistent with the original composite feature tensor is obtained.

6. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 5, characterized in that, The conditional implicit coding compactly characterizes the evolution of multi-dimensional signal primitives in the current composite feature tensor over time.

7. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 5, characterized in that, Based on the pseudo-deception feature tensor and the original composite feature tensor, discriminative deep features are extracted through the forward adversarial discrimination of the discriminator subnetwork to obtain discriminative deep features; Based on the discriminative depth features, a deception probability score sequence for the current navigation signal is calculated, including: Based on the pseudo-deception feature tensor and the original composite feature tensor, the discriminator sub-network concatenates the features of the two in the channel dimension to obtain the combined feature tensor. Based on the combined feature tensor, the discriminator subnetwork calculates attention weights along the time dimension and feature channel dimension through multiple parallel spatiotemporal attention units, and aggregates the features in each dimension according to the attention weights to obtain preliminary deep features. The preliminary deep features obtained by each spatiotemporal attention unit are then added and fused element by element to obtain the fused deep features. Based on the fused depth features, the discriminator subnetwork performs a global average pooling operation along the feature channel dimension to compress the channel information while retaining the complete length of the time dimension, thus obtaining the discriminative depth features, which are two-dimensional tensors. Based on the discriminative deep features, the discriminator subnetwork inputs the feature vector corresponding to each time step in the discriminative deep features into a fully connected layer with shared weights. This fully connected layer has a single output node and undergoes a time-step nonlinear mapping through the Sigmoid activation function to obtain a series of deception probability scores between 0 and 1. The scores of all time steps are arranged in chronological order to form a deception probability score sequence.

8. The method for detecting UAV satellite navigation spoofing interference based on adversarial neural networks according to claim 7, characterized in that, The deception probability score sequence is compared with a preset detection threshold. If the deception probability score is higher than the detection threshold, it is determined that there is a progressive synchronous deception interference, and an alarm signal is generated, including: Based on the deception probability score sequence calculated from multiple consecutive time windows, a moving average filter is used to smooth the deception probability score sequence to obtain a smoothed score value sequence. Each score in the smoothed score value sequence is compared with a preset detection threshold. If the number of scores in the smoothed score value sequence that continuously exceed the detection threshold reaches a preset confirmation threshold, it is determined that there is a progressive synchronization deception interference in the current navigation signal. In response to the determination that there is a progressive synchronous deception interference, an alarm signal is obtained, which includes the interference type identifier and the interference start time information.

9. A UAV satellite navigation spoofing and interference detection system based on adversarial neural networks, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The feature extraction module is used to collect satellite navigation baseband signals in real time in the multipath reflection micro-environment of urban canyons where the ratio of building height to street width is greater than 2, for satellite navigation receiver terminals mounted on the belly of UAVs with the antenna phase center located above the fuselage structure. The module extracts multi-dimensional signal primitive features of the satellite navigation baseband signals, performs zero-mean unit variance normalization and nonlinear dimensionality reduction mapping on each dimension of the signal primitive features to obtain the corresponding feature vectors, and aligns each feature vector along the time axis by sliding window and splices it with multi-dimensional tensors to form the original feature set. The fusion module is used to calculate the mutual information and temporal dependency strength between any two feature vectors in the original feature set, and construct a feature association graph with a fully connected topology. A multi-head attention mechanism is used to perform multi-level aggregation and updating of the node features in the feature association graph to obtain a composite feature tensor that integrates the information of the coordinated changes of each signal primitive in time step and spatial channel. The deception detection module is used to input a composite feature tensor into a pre-trained adversarial neural network, which includes a generator subnetwork and a discriminator subnetwork. The generator subnetwork simulates a continuous, gradually shifting sequence of progressive synchronous deception interference to obtain a pseudo-deception feature tensor. Based on the pseudo-deception feature tensor and the original composite feature tensor, discriminative deep features are extracted through forward adversarial discrimination by the discriminator subnetwork. Based on the discriminative deep features, a deception probability score sequence for the current navigation signal is calculated. The alarm generation module compares the deception probability score sequence with a preset detection threshold. If the deception probability score is higher than the detection threshold, it determines that there is progressive synchronous deception interference and generates an alarm signal.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images